How AI visibility is measured — and what a score actually means

AI visibility is measured two different ways that are constantly confused: a readiness score, which checks whether your store is technically capable of being read and cited, and a citation rate, which observes how often assistants actually name your store when asked real buying questions. Readiness is a deterministic checklist you can compute in seconds; citation rate requires repeatedly asking the assistants and recording what came back.

Every tool in this category shows you a number out of 100. Almost none of them tell you what the number is a measurement of. That is not a small omission — the two things a score could plausibly mean are so different that acting on one while believing the other will send you in the wrong direction.

This article separates them. It covers what a readiness score can and cannot tell you, how a citation rate is actually collected, why the sample size is inherently small and what to do about that, and what an honest measurement looks like when you build one yourself.

Two numbers wearing the same name

Start with the distinction, because everything downstream depends on it.

Readiness is necessary but not sufficient. You can pass every technical check and still never be cited, because being readable is a precondition for being recommended, not a cause of it. Conversely a citation rate with no readiness diagnosis tells you that you are losing without telling you why.

What a readiness score should actually check

A readiness score is only as good as its checklist, and the checklist should be published. If you cannot see which checks ran and how each was weighted, the score is unfalsifiable and you should not act on it.

The checks that carry real weight, roughly in order of how badly failing them hurts:

1. Crawler access — whether GPTBot, ClaudeBot, PerplexityBot and Google-Extended are permitted in robots.txt, and whether an edge or hosting rule blocks them by user agent before the request reaches your store. This is binary and it gates everything else.
2. Product content depth — whether descriptions answer identifiable buyer questions, or are two-line stubs with a size and a colour. This is the strongest predictor of being cited once you are readable.
3. Structured data — whether schema.org Product markup is present and complete, with price, priceCurrency, availability and brand populated on every product template variant.
4. Store guide — whether llms.txt or agents.md exists and describes your specific catalogue rather than serving a generic platform default.
5. Crawlability basics — sitemap present and current, canonical URLs consistent, no login wall or aggressive bot challenge on product pages.
6. Answerable non-product pages — shipping, returns, sizing and materials pages that state policies in plain sentences rather than in an accordion rendered only after JavaScript execution.

Notice what is absent: keyword density, word count targets, and anything resembling a domain authority figure. Those belong to a different discipline and importing them into a readiness score adds noise dressed as rigour.

How a citation rate is actually collected

A citation rate is collected by probing. You define a panel of buying questions, ask each assistant each question on a schedule, and record whether your domain appears in the answer or its citations. The mechanics matter more than they sound, because most of the ways this goes wrong are in the setup rather than the maths.

1. 1

   Write the question panel from buyer language, not brand language

   Ten to twenty questions phrased the way a customer would ask them: "best unscented shampoo for a sensitive scalp under $30", not "is [your brand] good". Brand-name questions almost always return your store and measure nothing. Derive the panel from your actual catalogue and your actual category.

2. 2

   Freeze the panel

   The panel must stay fixed between runs. If you change the questions, you have changed the instrument, and the before-and-after comparison is void. Add new questions as a separate series rather than editing the existing one.

3. 3

   Probe every engine you care about, on a schedule

   Ask each question of each assistant on the same cadence — weekly is a reasonable default. Record the full answer text and the cited URLs, not just a yes/no. You will want the raw text later when a result surprises you.

4. 4

   Record presence, position and competitors

   Three fields per probe: did your domain appear, where in the answer, and which stores were named instead. The competitor list is often the most useful output — it tells you who is currently winning the question you want.

5. 5

   Report the trend, never the single run

   Assistant answers vary between identical runs. One probe is an anecdote. The metric is the proportion of panel questions returning your domain, tracked over weeks, with the raw runs kept so any number can be traced back to the answer that produced it.

The sample size problem, stated honestly

Here is the part most vendors leave out. A citation-rate panel is small. Twenty questions across four assistants is eighty observations per week. That is enough to see a real change from zero citations to consistent citations. It is not enough to distinguish 35% from 40%, and any tool presenting a citation rate to one decimal place is implying a precision the method cannot deliver.

The practical consequence: treat citation rate as a coarse instrument. It reliably answers "are we in the conversation at all", "did that fix move us", and "who keeps beating us". It does not reliably answer "did we improve two points this week". Design your reporting around the questions it can answer.

The same discipline applies to how the number is computed. Use a median across runs rather than a mean where you aggregate — a single unusual answer should not drag the figure — and never smooth a series so aggressively that a real drop becomes invisible.

Attribution: the third number, and the most abused

Beyond readiness and citation rate sits the question everyone actually cares about: how much revenue came from AI. This is where honesty is hardest, because a confident-looking figure is easy to produce and almost always wrong.

Three tiers of evidence exist, and a measurement that does not distinguish them is misleading:

The last two rows are why every AI-revenue number is a floor. A shopper asks an assistant on Monday, thinks about it, and types your domain directly on Wednesday. That order is AI-influenced and permanently unattributable. Any figure that quietly folds guesses into the total to make it look bigger has stopped being a measurement.

Putting the three together

Each number has one job. Readiness tells you what to fix. Citation rate tells you whether the fix changed how assistants answer. Attributed revenue tells you whether that change reached the bank — as a floor.

The failure mode is using one to answer another's question. Readiness improving while citation rate stays flat means you became eligible without becoming preferable — the next work is content and third-party credibility, not more technical checks. Citation rate improving while attributed revenue stays flat may mean the questions you win are not commercially valuable, or may simply mean attribution is invisible for those surfaces. Distinguishing those two requires looking at which questions moved, not at the totals.

This is the loop Nomu runs — a deterministic readiness audit, a weekly probe panel per engine, and attribution that labels every order as confirmed, high-confidence or unknown rather than blending them. All the arithmetic is computed in code and shown behind "Show the math", because a number you cannot audit is a number you cannot defend to whoever asks you about it.

Frequently asked questions

The verdict

AI visibility measurement is young enough that the vocabulary has not settled, and that ambiguity is being used commercially. The defence is simple: for any number you are shown, ask what was observed to produce it. If the answer is "we checked your site", it is readiness. If the answer is "we asked the assistants and recorded what they said", it is a citation rate. If nobody can tell you, it is a marketing asset.

Measure both, keep the raw runs, report the trend rather than the point, and label attributed revenue as a floor. That is unglamorous and it is the only version of this that survives contact with someone who asks how the number was made.